Binarized Neural Networks (BNN) have recently been proposed as an energy-efficient alternative to more traditional learning networks. Here we study the problem of formally verifying BNNs by reducing it to a corresponding hardware verification problem. The main step in this reduction is based on factoring computations among neurons within a hidden layer of the BNN in order to make the BNN verification problem more scalable in practice. The main contributions of this paper include results on the NP-hardness and hardness of PTAS approximability of this essential optimization and factoring step, and we design polynomial-time search heuristics for generating approximate factoring solutions. With these techniques we are able to scale the verification problem to moderately-sized BNNs for embedded devices with thousands of neurons and inputs.
CITATION STYLE
Cheng, C. H., Nührenberg, G., Huang, C. H., & Ruess, H. (2018). Verification of Binarized Neural Networks via Inter-neuron Factoring: (Short Paper). In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11294 LNCS, pp. 279–290). Springer Verlag. https://doi.org/10.1007/978-3-030-03592-1_16
Mendeley helps you to discover research relevant for your work.